|Publication number||US8117142 B2|
|Application number||US 12/245,802|
|Publication date||Feb 14, 2012|
|Filing date||Oct 6, 2008|
|Priority date||Oct 4, 2007|
|Also published as||US20090094180|
|Publication number||12245802, 245802, US 8117142 B2, US 8117142B2, US-B2-8117142, US8117142 B2, US8117142B2|
|Original Assignee||Siemens Medical Solutions Usa, Inc.|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (11), Non-Patent Citations (5), Referenced by (1), Classifications (11), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
This application claims benefit under 35 U.S.C. §119(e) to U.S. Provisional Patent Application Ser. No. 60/977,451 filed on Oct. 4, 2007.
1. Field of the Invention
The present invention generally relates to nuclear medicine, and to systems for obtaining nuclear medicine images. In particular, the present invention relates to a method of real-time crystal peak tracking for positron emission tomography (PET) avalanche-photodiodes (APD) detectors.
2. Description of the Related Art
Avalanche-photodiodes (APD) detectors on positron emission tomography (PET) scanners suffer significant gain drifting due to thermal variations. The gain drifting causes crystal peak shifting in position profile. These disadvantages are not admitted to have been known in the art by inclusion in this section. However, for the foregoing reasons, there is a need for a method to compensate for gain drifting due to thermal variations.
The present invention is directed to a method of real-time crystal peak tracking for APD detectors that satisfies the need to compensate for the significant gain drifting due to thermal variations in APD detectors on PET scanners. In particular, the present invention utilizes the neural network based self-organized feature map (SOFM), and increases the tolerance of APD detectors to the thermal changes.
These and other features, aspects, and advantages of the present invention will become better understood with reference to the following description and appended claims, and accompanying drawings where:
The figures illustrate diagrams of the functional blocks of various embodiments. The functional blocks are not necessarily indicative of the division between hardware circuitry. Thus, for example, one or more of the functional blocks (e.g., processors or memories) may be implemented in a single piece of hardware (e.g., a general purpose signal processor or a block or random access memory, hard disk or the like). Similarly, the programs may be stand alone programs, may be incorporated as subroutines in an operating system, may be functions in an installed imagining software package, and the like.
It should be understood that the various embodiments are not limited to the arrangements and instrumentality shown in the drawings.
The present invention may be understood more readily by reference to the following detailed description of preferred embodiments of the invention as well as to the examples included therein.
All numeric values are herein assumed to be modified by the term “about,” whether or not explicitly indicated. The term “about” generally refers to a range of numbers that one of skill in the art would consider equivalent to the recited value (i.e., having the same function or result). In many instances, the term “about” may include numbers that are rounded to the nearest significant figure.
APD detectors on PET scanners suffer significant gain drifting due to thermal variations. The gain drifting causes crystal peak shifting in position profile. The present invention relates to methods of real-time crystal peak tracking for APD detectors, which preferably compensate for the thermal variations. Methods according to the present invention preferably increase the tolerance of APD detectors to thermal changes. Methods according to the present invention preferably make extended usage of the neural network based self-organizing feature map (SOFM) for building crystal lookup tables on PET detectors coupled with photomultiplier tubes (PMTs) and Avalanche-photodiodes (APDs).
A block position profile from which a crystal lookup table (CLT) is built contains millions of events, randomly detected by all pixels in a detector block. Therefore, use of a supervised neural network approach is not practical; however, a self-organizing (unsupervised) neural network scheme can solve the problem effectively. In particular, as illustrated in
In particular, the SOFM first determines the “winning” neuron a from n using the competitive layer equation
where p is a 2×1 input vector from the detector position profile and W is a weight matrix representing neuron positions. For an exemplary detector, e.g., a 20×20 block detector as used in the Siemens INVEON Dedicated PET scanner, there are 400 neurons representing the locations of each of the “pixels” of the detector block; accordingly, W is a 400×2 matrix (400 pixels, with X and Y positions of each). Vector a is the output from the competitive layer indicating the winning neuron. In the normalized case, Wp can be regarded as the distance between p and neurons weight vectors W. Vector a is the output from the competitive layer indicating the winning neuron. Only the neuron with the closest weight vector to p “wins” the competition each time.
Next, the weight vectors for all neurons within a certain neighborhood (e.g., a neighborhood of 1) for the winning neuron are updated using the Kohonen rule,
where α is a learning rate, q is the training index number, and w is the neuron in the neighborhood of the winning neuron (which has the same dimension of p).
When a vector p is presented, the weight of the winning neuron will move toward p. As a result, after many presentations of p from the position profile, the neuron will have learned the vectors that cluster together, which indicates a pixel location in the position profile. After training, the positions of the neurons in the weight matrix will represent the positions of the detector pixels. Thereafter, the crystal lookup table can be constructed easily from the pixel locations represented by the neurons.
By way of example of implementation of the method of the invention, the SOFM training may by simulated by a MathLab program. The 400 (exemplary) neurons are evenly (randomly) distributed in the area of interest as initial weight vectors.
The exemplary trained network is illustrated in
Thus, the developed neural network-based algorithm for CLT building produces higher pixel identification accuracy. There are two issues to be addressed in making this algorithm practical on a PET scanner. First, like all competitive networks, dead neuron and stability problems may occur, as apparent in the upper right-hand portion of
Second, the relatively lengthy time to train the system can be overcome by implementing the SOFM algorithm on a field programmable gate array (FPGA) chip within the event processing module (EPM) electronics to accelerate the training process. That makes it possible to build the CLT on-line while the position profile is being acquired and makes the CLT building time comparable to that of the prior art, but with significantly higher accuracy. A schematic illustration of such a system is shown in
The EPM electronics related to the SOFM training includes a large-scale FPGA chip and several Static Random Access Memory (SRAM) chips. The FPGA implementation consists of a front end module, a SRAM addressing module, a competition pipeline, and a neuron-training module. The total logic utilizes less than 20% of the FPGA resource.
A position profile is acquired first with a short period of time. Initial weight vectors of the exemplary 400 neurons based on the position profile are calculated and downloaded to the SRAM before training. The initial neurons are evenly (randomly) distributed in the interested area of the position profile as shown in
A lookup table is also generated based on the initial weight vectors of the neurons and stored in the SRAM. The lookup table defines regions indexed by the X, Y value of the event. Only the neurons in the same region (rather than all 400 neurons) compete with each other. The adjacent regions overlap each other to ensure that all necessary neurons are picked up for competition when an event comes in. The lookup table reduces the size of the competitive layer and the pipeline processing time, so the required FPGA resource and processing cycle does not increase with the number of neurons.
When an event triggers, the digitized ABCD values are sampled in the front end and the X, Y position of the event is calculated. The X, Y value is used as an address to retrieve the region number from the lookup table in the SRAM. The positions of all the neurons in the region are read from the SRAM and sent to the competition pipeline sequentially. The FPGA contains multiple 18 bit×18 bit multiplier blocks. Due to the pipeline implementation, only one multiplier is needed. The winner of the competition is the neuron to be updated by the X, Y value of the incoming event using the Kohonen rule. Each time an event happens, only one neuron is updated. The updated neuron is then stored back to the SRAM, and the training logic is ready for the next event. The neurons' positions are uploaded from the SRAM after the training is completed. A CLT is built based on the peak positions represented by the neurons' positions.
As has been observed in simulation, dead neuron and stability problem may occur. In particular, as shown in the Figures, one neuron is oscillating between two pixels while another neuron never wins the competition and is never trained by any incoming events. In FPGA training, some neurons are pulled toward the edge of the pixels due to the noise events in the gap. This implicates the signal-noise ratio that can be tolerated by the algorithm; applying a weighted learning rate based on histogramming can help alleviate such issues. Additionally, using a larger neighborhood size at the beginning of the training may also overcome the dead neuron issue.
Once the neuron is trained to the correct position, it is “trapped” around this location in future training unless the pixel is changed or the tube gains are unbalanced. This feature makes the FPGA implementation a good method for detector on-line monitoring. The logic can be run in parallel with the acquisition logic. When a detector defect happens due to a pixel cracking or tube drifting, it can be detected immediately rather than after image reconstruction.
The SOFM for building crystal lookup table on PET detector coupled with photomultiplier tubes (PMTs), as described above, is verified here with APD detector that is featured as low peak-background ratio. Under the neighborhood 1 rule, only the winning neuron is updated using the Kohonen rule,
where α is learning rate, q is training index number, p is position vector of incoming event, w is the winning neuron, which has the same dimension of p. As discussed above, this method works well with the Siemens INVEON detector. However, on APD detector, some neurons are pulled toward the edge of the pixels by the noise events in the peak gap, due to the low peak-background ratio. According to the present invention a weighted learning rate based on histogramming count is preferably applied to solve this issue. The Kohonen rule is preferably modified as
where C is the histogramming count at p position. The initial neurons and SOFM training results on an APD detector are shown in
According to the present invention, once the neuron is trained to the correct position, it is preferably trapped around this position in the further training. The algorithm is compact and simple, and can be easily implemented in the FPGA chip, preferably running in parallel to event acquisition. These features of SOFM make it a preferred method for detector real-time crystal peak tracking, according to the present invention. When the APD gains are drifting slowly, the neurons are capable of tracking the crystal positions.
The real-time crystal peak tracking feature of SOFM is verified through simulation. The left-down corner of the position profile in
Although the present invention has been described in considerable detail with reference to certain preferred versions thereof, other versions are possible. Therefore, the spirit and scope of the appended claims should not be limited to the description of the preferred versions contained herein.
The reader's attention is directed to all papers and documents which are filed concurrently with this specification and which are open to public inspection with this specification, and the contents of all such papers and documents are incorporated herein by reference.
All the features disclosed in this specification (including any accompanying claims, abstract, and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise. Thus, unless expressly stated otherwise, each feature disclosed is one example only of a generic series of equivalent or similar features.
Any element in a claim that does not explicitly state “means for” performing a specified function, or “step for” performing a specific function, is not to be interpreted as a “means” or “step” clause as specified in 35 U.S.C §112, sixth paragraph. In particular, the use of “step of” in the claims herein is not intended to invoke the provisions of 35 U.S.C §112, sixth paragraph.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US3798370||Apr 17, 1972||Mar 19, 1974||Elographics Inc||Electrographic sensor for determining planar coordinates|
|US5519224 *||Nov 30, 1994||May 21, 1996||Siemens Medical Systems, Inc.||Method of localizing scintillation events detected by scintillation camera systems|
|US6307203||Nov 25, 1998||Oct 23, 2001||Ge Medical Systems Global Technology Company Llc||Data binning method and apparatus for PET tomography|
|US6337481||Nov 15, 1999||Jan 8, 2002||General Electric Company||Data binning method and apparatus for pet tomography including remote services over a network|
|US7301144 *||Dec 29, 2004||Nov 27, 2007||General Electric Company||Method and system for calibrating a positron emission tomography system|
|US7515304||Jan 9, 2006||Apr 7, 2009||Seiko Epson Corporation||Smoothing of dispositions of lattice points for creation of profile|
|US7827129 *||May 18, 2007||Nov 2, 2010||Siemens Medical Solutions Usa, Inc.||Crystal lookup table generation using neural network-based algorithm|
|US20030055799 *||Jun 18, 2002||Mar 20, 2003||Starzyk Janusz A.||Self-organizing data driven learning hardware with local interconnections|
|US20050151084 *||Aug 10, 2004||Jul 14, 2005||Michael Zibulevsky||Gamma camera for emission tomography and method for adaptive event position estimation|
|US20060138315 *||Dec 29, 2004||Jun 29, 2006||General Electric Company||Method and system for calibrating a positron emission tomography system|
|US20060209066||Mar 15, 2006||Sep 21, 2006||Matsushita Electric Industrial Co., Ltd.||Three-dimensional image communication terminal and projection-type three-dimensional image display apparatus|
|1||Demuth, Howard and Beale, Mark, "Neural Network Toolbox for Use with MathLab-User's Guide, Version 3.0", Jan. 1998, pp. 7-1 to 7-34, The Mathworks, Inc.|
|2||Demuth, Howard and Beale, Mark, "Neural Network Toolbox for Use with MathLab—User's Guide, Version 3.0", Jan. 1998, pp. 7-1 to 7-34, The Mathworks, Inc.|
|3||Honkela, Timo, "Description of Kohonen's Self-Organizing Map", excerpt from "Self-Organizing Maps in Natural Language Processing", Jan. 1998, pp. 1-8.|
|4||Honkela, Timo, "Self-Organizing Maps in Natural Language Processing", Dec. 2007, pp. 1-63.|
|5||Jeong et al., "Performance Improvement of Small Gamma Camera Using Nal(TI)Plate and Position Sensitive Photo-Multiplier Tubes", Phys. Med. Biol., 2004, pp. 4961-4970, vol. 49, IOP Publishing Ltd.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US9064305||Aug 13, 2013||Jun 23, 2015||Samsung Electronics Co., Ltd.||Methods and apparatuses for generating a response of a scanner in an imaging apparatus and medical image using the same|
|U.S. Classification||706/16, 706/12, 706/17, 706/15, 706/18|
|International Classification||G06F15/18, G06E1/00, G06E3/00, G06G7/00|
|Nov 14, 2008||AS||Assignment|
Owner name: SIEMENS MEDICAL SOLUTIONS USA, INC., PENNSYLVANIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:HU, DONGMING;REEL/FRAME:021833/0422
Effective date: 20081113
|Jul 15, 2015||FPAY||Fee payment|
Year of fee payment: 4